Marketing Attribution & ROI Simulator (Shiny)

An interactive Shiny app for multi-touch attribution (Markov) and a lightweight MMM-style ROI scenario simulator (MMM-lite)—with Diagnostics and Data Health checks to make results easier to trust and explain.

🔗 Live Demo

Try the app here: https://ofosuosei.shinyapps.io/marketing-attribution/

🧭 What this app does

Upload & Validate

• Upload a CSV (or use included demo datasets)

Auto-detects dataset type:

• Journey (multi-touch events) → Markov attribution + journey diagnostics
• MMM (media spend / activity + sales) → MMM-lite scenario simulation + MMM diagnostics
• Preview the data and confirm required columns

Attribution (Markov)

• Computes removal-effect attribution (normalized)

Optional:

• Lookback window (days)
• Remove consecutive duplicate channels

Shows:

• Attribution chart + table
• Sample “top paths”

ROI Simulator (MMM-lite)

• Fit a simple MMM-lite model (per division if applicable)
• Adjust channel multipliers (sliders) to run “what-if” scenarios

Outputs:

• Baseline vs scenario predicted sales plot
• Scenario summary + per-period table (baseline, scenario, lift)

📈 Diagnostics

• MMM diagnostics
• Holdout metrics (train/test)
• Coefficients table
• Actual vs predicted (holdout)
• Residuals vs fitted
• Markov diagnostics
• Journey stats (users, touches, conversion rate, etc.)
• Top transitions
• Optional “health” warnings (e.g., suspiciously high conversion rate)

📊 Data Health

A credibility-focused panel to quickly spot common issues:

• Missingness by column
• Date/time coverage (min/max timestamps)
• Touches-per-user distribution (journey datasets)
• A configurable conversion-rate threshold (numeric input) to flag “suspiciously high” conversion rates

⚙️ Setup

Install packages

    install.packages(c("shiny","dplyr","ggplot2","DT"))

▶️ Run

From project root:

    shiny::runApp()
    # or
    shiny::runApp(".")

⚠️ Notes & disclaimers

MMM-lite is intentionally simple (useful for demos and directional “what-if” scenarios). It is not a full Bayesian MMM, does not automatically handle adstock/saturation unless explicitly implemented in your MMM script, and should not be treated as production-grade causal inference.

Attribution results depend heavily on data definitions (especially what counts as a “conversion” and how journeys are filtered).